Spaceballs the Datacenter!
Why Musk's xAI-SpaceX Datacenter in Space is Bullshit
I Did the Math on Musk’s Space Datacenter. It’s Dumber Than Mega Maid.
Read the full analysis on my website: “Spaceballs the Datacenter!”
Remember Spaceballs? Mel Brooks’ 1987 masterpiece where the villains try to vacuum Druidia’s atmosphere with a giant space maid?
Musk’s plan to merge xAI with SpaceX and build datacenters in orbit makes less sense than that plot.
I did the math. Eight hours of Stefan-Boltzmann equations, thermal radiator sizing, and reaction wheel calculations. The numbers are brutal.
One GPU in space requires a satellite the size of a minivan. 287 kilograms. $4,700,000 at launch. And that’s being extremely generous to Elon.
The problem that nobody in the breathless press coverage seems to understand is that space has no air. No convection. The only way to dump heat is radiation. And radiation follows physics that really, really hate you.
Every NVIDIA B200 pumps out 1,000 watts of heat. To radiate that heat away, you need 2.91 square meters of radiator panel. That’s a regulation beer pong table. Per GPU. In orbit. Just to keep things from melting.
But wait. You also need solar panels to power the GPU. Those need to charge batteries for the 35 minutes of every 92-minute orbit when you’re in Earth’s shadow. So add another 5.3 square meters. Now you need attitude control to point all those panels correctly. And structure to hold it together. And communications to talk to the ground.
One GPU. 287 kg. 13 m2 fully deployed.
“Economies of scale!” you say. “Put MORE GPUs on each satellite!”
I tried that. It gets ugly.
At 32 GPUs, your satellite has so much deployed area that solar radiation pressure and atmospheric drag start spinning it like a rotisserie chicken. Your cute little reaction wheels can’t keep up. You need Control Moment Gyroscopes. The kind the ISS uses. They weigh 128 kg. *Each.* You need four.
By the time you hit 64 GPUs, you’re building a satellite that weighs more than the Hubble Space Telescope. For one rack’s worth of computing.
To match a single terrestrial datacenter with 100,000 GPUs, you’d need 3,449 of these satellites each with 29 GPUs (I call them BallSats). Fifteen million kilograms to orbit. 247 Falcon Heavy launches. At SpaceX’s current pace, that’s a year and a half of launches dedicated to nothing but GPU satellites.
Total cost: roughly $98 billion per Spaceballs Datacenter.
A terrestrial datacenter with the same capacity? $5 billion. Including the building, cooling towers, and a cafeteria with a salad bar.
Musk’s version costs 20x more. And it’s worse at everything except “having a nice view.”
I haven’t even mentioned latency. Distributed GPU training needs sub-microsecond synchronization. Space gives you 20-50 milliseconds. That’s 10,000 times too slow. The speed of light and tyranny of distance kills this application before the economics do.
I wrote the full breakdown with all the physics, all the Spaceballs gifs, and all the snark. It’s about 10,000 words. Tables. Equations. A surprisingly moving story about a guy who folds origami at MIT bars… Read the full analysis here: “Spaceballs the Datacenter!”
You can’t engineer your way out of thermodynamics. You can’t MBA your way past Stefan-Boltzmann. And you definitely can’t hype your way through the speed of light.
May the Schwartz be with you.
BY THE WAY! If you’re into first-principles forecasting, which is what this project is, I’m leading a workshop on Applied Forecasting at RSAC 2026 in San Francisco. Details here.